For the first time this year, SpaceXAI (formerly xAI) competes at the AI frontier again. Although Anthropic’s Claude Fable 5 is out of reach and GPT-5.6 will likely surpass its results, Grok 4.5 sits roughly at the level of Opus 4.8 and GPT-5.5 in overall scoring. More importantly, it excels at coding and beats out GLM-5.2’s cost per task significantly. One caveat if you’re in the EU: it won’t be available until mid-July.
Grok 4.5 is the first LLM launched after the SpaceXAI rebrand earlier this week. It is also the first to benefit from Elon Musk’s 60 billion dollar purchase of coding agent provider Cursor. Oddly, there are two separate launch blogs, one each from SpaceXAI and Cursor. The benchmarks provided overlap but aren’t quite the same ones. Regardless, the independent firm ArtificialAnalysis puts Grok 4.5 fourth in line in terms of overall intelligence, reaching a Sonnet 5-beating 54 compared to GPT-5.5 at its highest reasoning step at 55, Opus 4.8 Max at 56, and Fable 5 at 60.
A swift return of the pricing wars?
This summer of frontier AI releases has jumped narratives several times already. Fable 5 kicked things off and remains clearly the strongest overall LLM, albeit at a fairly staggering 25 dollars per 1 million output tokens. GLM-5.2, an open-weight challenger from China’s Z.ai, undercuts this enormously at $4.40/1M tokens. However, it is incredibly inefficient, reasoning for a very long time indeed. Grok 4.5, at $6/1M output tokens if you stay within 200k input tokens, will be even cheaper in practice. Per SWE Bench Pro task, Grok 4.5 at a blistering 80 tokens per second consumes over four times fewer tokens than Opus 4.8 at max reasoning for an equivalent score. Quicker outputs, lower costs, pretty much the same result.
Musk calls Grok 4.5 “Opus-class” for good reason then. The question begged here is whether or not we’ve just restarted the race to the bottom in terms of pricing. Just when Fable 5 sets a new high floor for costs, organizations are fretting about spiralling token spend. Grok 4.5 seems like a real answer here, and it’s likely that anyone with a frictionless model choice could switch rather quickly from GPT-5.5 or Opus 4.8 to this far cheaper equivalent from SpaceXAI. However, that doesn’t quite tell the whole story.
Grok 4.5 is ultimately not part of the new breed of LLMs emerging from OpenAI and Anthropic. It seems to us that Fable 5 has a tier all to itself, with the upcoming GPT-5.6 still needing to prove it’s up to that level. Irrespective of that development, though, SpaceXAI can now serve as a legitimate enterprise solution. The Cursor integration helps tremendously here, as it’s a coding agent that sits third behind Claude Code and GitHub Copilot in overall adoption. With high costs for the two leading firms (and GitHub Copilot overall seeming to stumble), Cursor has room to grow and compete more.
The EU sticking point
We’re a long way away from Grok’s initial pitch of being the “maximally truth-seeking” and “based” LLM that has sparked controversy on X for the past few years. Under investigation via the EU’s Digital Services Act, X can thank this level of scrutiny to its alleged lack of proper risk assessment and mitigation for Grok’s potential dangers. This has all meant that for the Grok line of models to serve as a serious enterprise solution, it needs to overcome untold reputational damage. Now, a delayed EU release suggests some degree of cooperation with the authorities here is taking place.
Grok 4.5 overall seems like a clear step in the right direction. It is very much focused on coding, which has resulted in a deeply impressive Artificial Analysis’ Coding Agent Index score. If we treat this as the right yardstick (and this index, along with AA’s overall scoring, tends to be close to the subjective overall sentiment around new LLMs), Grok 4.5 is within a whisker of the overall lead. Fable 5 in Claude Code scores a 77, GPT-5.5 in Codex a 76, and Grok 4.5 in Grok Build reaches that same score.
Cursor, it appears, has already contributed to this immensely. Trained on the most difficult tasks that the coding agent developer has faced, a deliberately broader training set than the earlier Composer 2.5 model has led to additional benefits. Although Grok 4.5 is specialized, it has general knowledge that seems to boost its capability.